Algebraic cycles from a computational point of view

Theoretical Computer Science 392 (2008) 128–140 www.elsevier.com/locate/tcs

Algebraic cycles from a computational point of view Carlos Simpson CNRS, Laboratoire J. A. Dieudonn´e UMR 6621, Universit´e de Nice-Sophia Antipolis, 06108 Nice, France

Abstract The Hodge conjecture implies decidability of the question whether a given topological cycle on a smooth projective variety over the field of algebraic complex numbers can be represented by an algebraic cycle. We discuss some details concerning this observation, and then propose that it suggests going on to actually implement an algorithmic search for algebraic representatives of classes which are known to be Hodge classes. c 2007 Elsevier B.V. All rights reserved.

Keywords: Algebraic cycle; Hodge cycle; Decidability; Computation

1. Introduction Consider the question of representability of a topological cycle by an algebraic cycle on a smooth complex projective variety. In this paper we point out that the Hodge conjecture would imply that this question is decidable for varieties defined over Q. This is intuitively well-known to specialists in algebraic cycles. It is interesting in the context of the present discussion, because on the one hand it shows a relationship between an important question in algebraic geometry and the logic of computation, and on the other hand it leads to the idea of implementing a computer search for algebraic cycles as we discuss in the last section. This kind of consideration is classical in many areas of geometry. For example, the question of deciding whether a curve defined over Z has any Z-valued points was Hilbert’s tenth problem, shown to be undecidable in [10,51, 39]. The same question for Q is still open but thought to be undecidable, and for other fields it is also an active research question [41]. Mazur has done a survey of related decidability questions in number theory and arithmetical geometry in [38], and Macintyre surveys the relationship between model theory and algebraic geometry in [37]. Decidability of various kinds of recognition problems for manifolds is a typical question in differential geometry, again known to be undecidable in dimension ≥ 4 because any group can be the fundamental group of a 4-manifold, and the word problem is undecidable in general. Nabutovsky has considered related questions [44,46] and has also used these notions to give geometric statements about sizes of geodesics [45] (I thank Y. Yomdin for pointing out these interesting references). Hales used computer approximation to differential-geometric questions in his proof of the Kepler conjecture [25], and computation is useful for enumerating complex ball quotients [49,48]. Hrushovski has considered many questions relating logic and geometry, for example he gives an algorithm to calculate the differential Galois group of an equation [29]. The problem of computing the topology of a manifold or submanifold from a collection of data points has recently been the subject of much investigation, see for example [7,47,56]. Similarly, the E-mail address: [email protected]. URL: http://math.unice.fr/∼carlos/. c 2007 Elsevier B.V. All rights reserved. 0304-3975/$ - see front matter doi:10.1016/j.tcs.2007.10.008

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

129

problem of approximating the Hodge theory of a manifold by finite element approximations goes back to Dodziuk [15] and again has been the subject of more recent works such as [21,24,27]. Our observation will be very elementary in comparison with all of these works. We work with smooth varieties defined over Q. The topology of such a variety may be computed algorithmically, either by triangulating [2] or by a point-cloud method [47] as we shall discuss in Section 2. Similarly, the Hodge filtration of the algebraic de Rham cohomology involves only calculating with algebraic numbers, and while the integrals of algebraic de Rham cohomology classes over explicit topological cycles are highly transcendental objects, these integrals can be approximated to arbitrary precision by algorithmic methods, Section 3. These two procedures come together when we consider the Hodge conjecture, and we do not need to consider non-algebraic real numbers in order to understand its formulation. If we assume the Hodge conjecture, then a topological cycle is either an algebraic cycle, in which case it can be found by a search over all subvarieties which may themselves be assumed to be defined over Q; or else it is not even a Hodge cycle, in which case the fact that it is not a Hodge cycle can be ascertained by showing that some Hodge integral does not vanish. Nonvanishing of an integral can be shown by a sufficient approximation. The computation of the topology of a real algebraic variety appears in [2,4,44,47,53,56], and the approximation of differential geometry using proven analysis by interval arithmetic was used by Fefferman and his co-workers [19,20,26,34], more recently see for example Mitrea–Mitrea [40]. If the basic idea was simple to explain, in Sections 2 and 3 some aspects of the proof will be reviewed in more detail, although our discussion will still be sketchy in many places. One of the goals of this discussion is to show that certain computations are in principle straightforward, in particular we prefer brute over more refined techniques if possible. This statement would be a good target for a computer-formalized proof, as it builds on a constructive point of view towards much of algebraic geometry, differential geometry and topology. The fact that the Hodge conjecture implies that a certain algorithm stops, could be taken as suggesting that it might not be true, since decidability is a very strong condition. On the other hand, the Lefschetz (1, 1)-theorem, which is the Hodge conjecture for codimension 1 cycles, gives decidability in this case, and it is hard to see any obvious reason why decidability should be different for codimension 1 or higher codimension. There may be interesting logical consequences of the Lefschetz (1, 1)-theorem which could be investigated. Even if the Hodge cycle condition turns out to be insufficient for characterizing algebraic cycles, one might ask whether some collection of Hodge-theoretic type invariants could characterize them. In this case, there would still be the analogue of the above result. A weakening of the Hodge conjecture would be that the problem of representability of topological cycles by algebraic cycles, should be decidable (Question 2.9). It is natural to try to implement the algorithm in specific cases. In Section 4 we discuss where one might start looking. It turns out that it is a nontrivial matter just to come up with a good variety X and a good topological cycle η ∈ H2 p (X, Q) for which to look for algebraic representatives. 2. Algebraic cycles Before getting to the statement about algebraic cycles, we need some preliminary considerations about computation of the topology of algebraic varieties. Some basic references are [2,4,28,32]. We are only giving an outline of the proof, and when discussing computability one should try to find well-behaved algorithms, so there is much room for improvement, see [1] for a number of recent references on this subject. For our purposes we only look at the existence of an algorithm, so we are undoubtedly ignoring important aspects for actual implementation. 2.1. Algebraic numbers We will be working with varieties defined over Q, which to be concrete is viewed as the ring of “algebraic complex numbers”. Use Q[i] as an approximation to the set of complex numbers. Elements should be regarded as simply points in Q2 ⊂ R2 = C. Definition 2.1. An algebraic complex number is a triple ( f, x, ε) where f ∈ Q[t], x ∈ Q[i], and ε ∈ Q with 0 < ε < 21 , such that: (a) f is irreducible; (b) | f (x)| < ε 2 | f 0 (x)|; and (c) for all y in the disc D(x, ε) we have | f 0 (y) − f 0 (x)| ≤ 14 | f 0 (x)|. Note that (a) and (b) are decidable statements whereas (c) will be obtained by a constructive bound on the second derivative.

130

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

One could replace this very basic definition by any of the several more advanced and algorithmically efficient techniques for isolating the real roots of a real algebraic polynomial [1,2,28,42]—to be applied to both the real and imaginary parts of a complex number. Given an irreducible polynomial f ∈ Q[t] we can determine its roots as algebraic complex numbers according to Definition 2.1. Looking at the sizes of the coefficients we obtain a bound |z| ≤ C1 for the roots plus a bound for the first and second derivatives in this region. Writing a f + b f 0 = 1 we can choose ε small enough so that if z is a root then for any x, y in the disc D(z, 2ε) we have | f 0 (x)| ≥ c2 and | f 0 (y) − f 0 (x)| ≤ c2 /4. A finite search approximates the roots sufficiently well which gives (b) and (c). If the conditions hold then there is a unique root of f in the disc D(x, ε) so that we can stop when we have found d = deg( f ) disjoint discs. Suppose we are given an algebraic complex number ( f, x, ε). Newton’s method gives a sequence {xi } approximating the unique root of f in D(x, ε). The number ε is called the accuracy of the algebraic complex number. Replacing x by some xi we can “increase the accuracy” to any smaller ε 0 < ε. Proposition 2.2. We can computably add, multiply and divide algebraic complex numbers. We can decide whether two algebraic complex numbers are “equal”: the polynomial f is the minimal polynomial of the root since it is irreducible, so two numbers are equal if and only if the respective discs overlap and the polynomials are rational multiples of one another. The set Q as we have defined it is enumerable. Proof. Adding, multiplying or dividing two elements x, y ∈ Q yields a new z, and from the minimal polynomials f and g we obtain a polynomial h ∈ Q[t] such that h(z) = 0. Factor h into products of powers of relatively prime irreducible factors, which can be done computationally. Locate approximately the roots of the factors as described above. The different factors do not share any common roots, so a brute approximation will eventually stop with all the roots being distinguished, i.e. the discs non-overlapping. Now approximate the original x and y sufficiently well so as to be able to tell which root of h should be z; take the appropriate irreducible factor of h and this gives an element z ∈ Q. The data of an element of Q as we have defined it is combinatorial so it is enumerable.  In what follows, denote by Q the set of algebraic complex numbers in the above sense, and use this as usual. The intersection Q ∩ R is decidable and enumerable too. It would be good to undertake all of our discussion below, for more general fields Υ ⊂ C. The key properties required of Υ are decidability of the equality and operations of Υ as in Proposition 2.2. The reader is invited to elaborate more precise statements in this direction. 2.2. Metric geometry on projective space Consider algebraic varieties as being embedded in projective space. It is necessary to manipulate distances on Pn . Using the Fubini-Study metric directly would get us into a discussion of how to compute geodesics. We bypass this by defining a distance function which is not geodesic but which will give the Fubini-Study metric on tangent spaces. Fix the standard hermitian form h·, ·i on Cn , also defined on Q[i]n . Suppose x = (x0 , . . . , xn ), y = (y0 , . . . , yn ) ∈ Cn+1 − {0}. These represent points denoted [x] and [y] in Pn . Put d([x], [y])2 := 1 −

|hx, yi|2 . |x|2 |y|2

This is independent of the choice of representatives for [x] and [y], and is computable if x, y ∈ Q[i]n+1 . For points close by it serves as a distance function. We can define in a similar way line segments between two nearby points, and standard simplices on a given set of vertices if the vertices are close together. 2.3. Bounding geometry Suppose X is a compact real algebraic manifold given by concrete equations (in other words, suppose it is a connected component of the set of real points of a complex algebraic variety defined over Q ∩ R). Suppose f : X → R is a polynomial function defined over Q. Then we can compute an upper bound for f on X by looking at the coefficients of the polynomials.

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

131

We usually apply this to functions on a complex projective manifold. For example, if f is a section of a line bundle OPn (k) then using the Fubini-Study metric on OPn (k) we can consider | f |2 as a real valued function, and bound it. Natural bundles such as the tangent or cotangent bundle or more generally ΩPi n or even jet bundles of natural vector bundles, may be considered as concrete quotients of direct sums of standard line bundles. We can choose an explicit real algebraic splitting, so we may similarly bound sections of bundles of differential forms. In particular, if f is a section of a line bundle then we may assume that its derivative and second derivative are globally bounded by a computable constant. 2.4. Smooth projective varieties A projective variety will mean the subset X ⊂ Pn defined by a finite collection of homogeneous equations f 1 , . . . , f m ∈ Q[z 0 , . . . , z n ]. We would like to have a computational meaning for the fact that X is smooth of dimension r . Thus, we define an explicit smooth projective variety of dimension r to be a collection ( f 1 , . . . , f m ; h 1 , . . . , h k ; I K ; P) where f i and h j are homogeneous equations such that the open subsets defined by h j 6= 0 cover Pn and each h j 6= 0 is contained in one of the standard affine opens for Pn . The f 1 , . . . , f m are the equations for X . Let x1 ( j), . . . , xn ( j) be the linear coordinates on this affine open, and let F1 ( j), . . . , Fm ( j) ∈ Q[x1 , . . . , xn ] be the dehomogenizations of the f i . In these terms, I K is the data for each j of two collections of indices i 1 ( j), . . . , i n−r ( j) ∈ 1, . . . , m and k1 ( j), . . . , kr ( j) ∈ 1, . . . , n such that on h j 6= 0 the ideal generated by the Fi p ( j) ( j) contains the other Fq ( j), and the ideal generated by the Fi p ( j) ( j) and their Hessian determinant with respect to the coordinates x` with ` different from one of the ku ( j), is the unit ideal in the localized ring Q[x]h j . In this case the coordinates xk1 ( j) , . . . , xkr ( j) provide local coordinates for X in the neighborhood h j 6= 0. And P is the collection of appropriate explicit ring elements needed for the proofs of these things, that is it contains the coefficients of all the Bezout-like expressions which are involved [6]. Given an explicit variety X as above, we can interpret the defining equations as sections of standard line bundles, and bound their derivatives and second derivatives. Cotangent vector bundles and the like are given explicit trivializations using the coordinates given by I K . The explicit ring elements given to say that the ideal generated by the functions and the Hessians is the unit ideal, provide a lower bound for the size of the Hessian determinant at any point of X . Combined with the upper bound for the second derivatives, for any  ∈ Q>0 we can calculate a number δ ∈ Q>0 such that on any ball of radius δ centered at a point of Pn (Q[i]), either one of the f i is bounded away from 0, or else the matrix of derivatives of the n − d defining equations given by the index I for the appropriate j is, when rescaled to a coordinate system of radius 1, -close to being constant with a determinant of 1 and norm bounded by a constant C calculable depending only on X . In this rescaled ball, the variety X looks up to an -small amount like a linear space with coordinates xk1 ( j) , . . . , xkd ( j) . 2.5. Computing the topology Suppose we are given an explicit smooth projective variety over Q as discussed above. Then we can compute its topology, for example X may be considered as a real semialgebraic set and the triangulation and computation of the homology is discussed in Basu–Pollack–Roy [2], also in [1,4,56]. As the question of triangulation can be a bit delicate, ˇ we discuss here an easier, cruder approach. The idea is to use the Cech complex for a covering of a tubular neighborhood by coordinate balls. This is related to the recent work by Niyogi, Smale, Weinberger, Carlsson and others [47,7, 8,17] on the recovery of the topology from random point-cloud data. I would like to thank Persi Diaconis for pointing out these references. Our discussion is entirely elementary, though: rather than using a random cloud of points, we take all points of a certain height [32] at a certain distance from the variety, that way we are sure to get the right topology. We can define abstractly the gradient flow for the function “distance to X ”. This abstract vector field is welldefined if not necessarily easy to compute. However, in any of our approximating balls as described in the previous paragraph, the abstract gradient vector field will differ from a standard one by an ε-small amount. Similarly, we can replace geodesics by pseudogeodesics which are segments of real algebraic curves defined by explicit equations. A pseudogeodesic between two nearby points well approximates the actual geodesic. Define the computable tubular neighborhood of X , with height N and radius the set of points in Pn which P ρ, to be 2 can be expressed as z = [z 0 : . . . : z n ] with z i ∈ Z[i] and |z i | ≤ N , such that i | f i (z)| ≤ ρ. Call this T (X ; ρ, N ). It results from the bounded geometry that there are computable constants c and C such that for any point z ∈ T (X ; ρ, N ) we have d(z, X )2 ≤ Cρ and conversely, for any point z of height N with d(z, X )2 ≤ cρ then z ∈ T (X ; ρ, N ).

132

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

Defined in this way, the neighborhood might be fairly elliptical, that is we cannot necessarily have c, C → 1. Define the puffed-up tubular neighborhood to be [ B(z, δ). P(X ; ρ, N , δ) := z∈T (X ;ρ,N )

Lemma 2.3. If δ is small enough, then if we choose ρ even much smaller, and finally choose N sufficiently large, then P = P(X ; ρ, N , δ) looks approximately like an honest tubular neighborhood of X of radius δ, in particular the boundary of P is piecewise spherical with the smooth locus having tangent plane close to the tangent plane of the actual tubular neighborhood. Therefore, the abstract gradient flow vector field provides a deformation retraction of P onto X .  We get from this lemma an explicit open subset of Pn , covered by a computable finite collection of coordinate balls (recall that we are using the easy version of the distance function), such that P is homotopy-equivalent to X . Furthermore, the coordinate balls are geodesically convex in Pn , so the multiple intersections in the covering are contractible. We can compute whether or not a finite collection of open coordinate balls, with coordinates in Q[i] and radii in Q, intersect or not. Therefore, we can compute the simplicial nerve of this covering. Recall that this is a simplicial set with a k-simplex for each nonempty k + 1-fold intersection of balls in the covering. This can be translated into a finite geometric simplicial complex in Pn in the following way. For k = 0, realize the vertices of the simplicial set by the centers of the balls in our covering. Then proceeding by induction on k = 1, 2, . . . , realize each nondegenerate simplex in the nerve as a geometric simplex in Pn by pseudo-geodesically contracting the boundary (which is already defined by induction) to one of the vertices. In this way, each simplex in our simplicial set is represented by a real-algebraic geometric simplex in Pn with equations defined over a fixed field Q[i]. This simplicial complex is located within the tubular neighborhood. One can then apply the approximate gradient flow for the distance to X , to make everything approach X . Proposition 2.4. The computable simplicial nerve of the covering of P defined above, is a simplicial complex H(X ) homotopy-equivalent to X . We can choose a subset of points in T (X ; ρ, N ) which still gives a covering, and for which all multiple intersections which are nonempty, also intersect X . By approximating the gradient flow of the distance to X , we can choose representatives for the simplices which get arbitrarily close to X in C 1 norm.  There are only finitely many nondegenerate simplices in our simplicial set, so we can compute a basis for the homology. The topological cycles in this basis are now represented by formal sums of geometric simplices in Pn , real algebraic over Q[i]. Corollary 2.5. Given an explicit smooth projective variety X , we can explicitly compute the Q-vector space Hi (X, Q). Furthermore, classes here can be given sequences of representatives which become arbitrarily close to X in C 1 norm.  One can define via explicit formulae the intersection product on the homology of the simplicial complex H(X ), and it will be the case for a smooth projective variety of dimension d that Hi (X, Q) satisfies Poincar´e duality of dimension 2d. The fact that this is true is a finite calculation on the simplicial complex. 2.6. Functoriality If X ⊂ Pn is an explicit smooth projective variety, then an explicit smooth subvariety is an explicit smooth projective variety Y given by equations g1 , . . . , gr such that the equations f 1 , . . . , f m of f are among the gi . Classical constructions such as the Veronese embedding of the product, remain explicit. An explicit affine open covering of X is a covering of X given by the complements of zero sets of a collection of sections h i ∈ Γ (X, O X ( p)) which generate O X ( p). On these affine open sets or their multiple intersections we can work computationally with functions, in particular we can define the notion of vector bundle. We can define the notion of explicit coherent sheaf on X , to be given by a matrix of homogeneous forms giving an exact sequence O X (−k)⊕a → O X (−l)⊕b → F → 0.

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

133

For an explicit vector bundle, looking at the order of poles of the transition functions, then doing the same on the kernel, gives a presentation. The bundles of differential forms Ω Xi are explicit vector bundles, as is the case for other similar constructions like jet bundles. 0 A morphism from X to Pn consists of a line bundle L on X (given by explicit transition functions) and sections u 0 , . . . , u n 0 ∈ Γ (X, L) defined by regular functions on the open covering where L is trivialized, together with local Bezout expressions saying that these sections generate L. On the affine open sets where L is trivial these sections give 0 0 0 maps to An +1 projecting to the map to Pn . Homogeneous polynomials on Pn pull back to sections of L ⊗k . Thus, 0 if Y ⊂ Pn is an explicit smooth projective variety then it makes sense to ask that the pullbacks to X of the defining equations of Y vanish on X , in other words locally are in the ideal of X . Thus, a morphism from X to Y consists of the data of a line bundle L, a family of sections u 0 , . . . , u n 0 ∈ Γ (X, L) which generate L, and local expressions of the pullbacks of the defining equations for Y , as vanishing on X in other words, being combinations of the functions defining X . Lemma 2.6. Suppose f : X → Y is an explicit morphism in the above sense. Then the graph of f is an explicit 0 subvariety G( f ) ⊂ X × Y ⊂ Pn × Pn ⊂ P N .  We can now define the map of homotopy types from X to Y , by looking at the diagram H(X ) ← H(G( f )) → H(Y ). Make the choices in such a way that we can define these maps as simplicial maps. We know that the leftward map induces an isomorphism on cohomology; given that this is known, the inverse to the isomorphism can be computed by just lifting basis elements. Thus we obtain a diagram ∼ =

Hi (X, Q) ← Hi (G( f ), Q) → Hi (Y, Q), with all the arrows and the inverse computable. This gives the morphism of functoriality Hi (Y, Q) → Hi (X, Q). 2.7. The algebraic cycle algorithm Definition 2.7. A class η ∈ H2 p (X, Q) is an algebraic cycle if there exists a morphism of explicit smooth projective varieties f : Y → X such that Y has dimension p, and η is in the image of the map H2 p (Y, Q) → H2 p (X, Q). This is equivalent to the usual notion of being an algebraic cycle. If η is represented by a complex algebraic variety Y 0 then by a standard specialization argument [55] Y 0 can be assumed to be defined over Q, and by the resolution of singularities it can be replaced with a smooth model Y . Proposition 2.8. Suppose X is an explicit smooth projective variety, and suppose η ∈ H2 p (X, Q). There is an algorithm to show that η is an algebraic cycle: if this algorithm stops successfully then η is an algebraic cycle and the algorithm will provide the map f : Y → X ; if in the sense of ZFC η is an algebraic cycle, then the algorithm will stop. Proof. The algorithm is to run through all possible explicit smooth projective varieties Y together with maps f : Y → X . The set of such is enumerable because Q is enumerable.  On the other hand, it is not at all clear a priori that there is an algorithm which stops exactly when η is not an algebraic cycle. Question 2.9. Is the problem whether a given class η ∈ H2 p (X, Q) is an algebraic cycle, decidable? We will see that the Hodge conjecture would imply that the answer to this question is “yes”.

134

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

3. Hodge cycles Recall that the algebraic de Rham cohomology of X is the hypercohomology of the complex d

d

d

Ω X· := O X → Ω X1 → · · · → Ω Xr . ˇ The hypercohomology H Di R (X ) := Hi (X, Ω X· ) may be computed as Cech cohomology with respect to any affine open covering. Define the Hodge filtration of subcomplexes by p d

d

F p Ω X· := Ω X → · · · → Ω Xr . Put   F p H Di R (X ) := Im Hi (X, F p Ω X· ) → Hi (X, Ω X· ) . This defines a decreasing filtration called the Hodge filtration of the de Rham cohomology group. There is an integration pairing Z : Hi (X, Q) × H Di R (X ) → C defined concretely in the following way. The procedure described by Bott and Tu in [5] associates to any de Rham cohomology class ϕ ∈ H Di R (X ) a differential i-form on X defined using a partition of unity with respect to the covering for which the hypercohomology was defined. Then integrate this i-form over the representative for a class η ∈ Hi (X, Q). 2p

Definition 3.1. A Hodge cycle is a homology class η ∈ H2 p (X, Q) such that for every ϕ ∈ F p+1 H D R (X ) we have R η ϕ = 0. The classical observation behind the theory of Hodge cycles is that an algebraic cycle is automatically a Hodge cycle. The reason is simple: a form ϕ ∈ F p+1 gives rise to a differential form which at every point has at least p + 1 factors of the form dz j , hence no more than p − 1 factors of the form dz j . Such a form restricts to zero on any p-dimensional algebraic subvariety of X , so it integrates to zero on any algebraic cycle. Recall the Hodge Conjecture. Every Hodge cycle is an algebraic cycle. Needless to say there is a massive amount of literature on this subject [35]. A more usual description of the statement is as follows. The Hodge filtration is the filtration associated to the Hodge decomposition M H Di R (X ) = H p,q (X ) p+q=i

and H p,q (X ) is the group of cohomology classes represented by currents of bidegree ( p, q). Identifying homology and cohomology by Poincar´e duality, the group of Hodge cycles may also be described as H 2 p (X, Q) ∩ H p, p (X ). Thus, the Hodge conjecture states that any closed current of bidegree ( p, p) with rational cohomology class, differs from a rational combination of algebraic cycles by a coboundary. To mention a couple of more recent formulations, Thomas rewrites the conjecture as an inductive statement about the existence of singular hyperplane sections [54], and Fabre replaces the notion of algebraic cycle by that of locally residual current of type ( p, p) [18]. 3.1. Computation of the Hodge filtration Continuing with the assumption that X is an explicit smooth projective variety over Q. The de Rham cohomology 2p and its Hodge filtration are defined over Q. In particular, the subspace F p+1 H D R (X ) has a basis consisting of elements defined over Q so in the definition of Hodge cycles it suffices to look at the integrals over forms defined over Q. 2p There are different possibilities for setting up an algorithmic computation of H D R (X ) with its filtration, based on making explicit the residue calculus originated by Griffiths, or the hypercohomology calculations of Grothendieck and Hartshorne. In what follows, we adopt the hypercohomology point of view, and suppose that an explicit affine open covering {Ui } of X has been chosen. We may further assume that the tangent bundle is trivialized on each open

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

135

2p ·,· ˇ set, and then write down the Cech double complex C D R (X ) for calculating H D R (X ). The components are isomorphic to direct sums of spaces of sections of the trivial bundle on multiple intersections of our affine covering, yielding an expression of the form M p,q C D R (X ) ∼ Γ (Ui0 ,...,iq , O X )a p = |I |=q+1 p

where a p is the rank of Ω X . The Hodge filtration comes from one of the filtrations on the double complex. An algebraic de Rham cohomology class in a certain level of the Hodge filtration may therefore be seen as a finite collection of sections in various Γ (Ui0 ,...,iq , O X )a p . In turn, these spaces of sections may be computed as quotients of polynomial algebras localized at a finite set of functions and divided by the ideal of X . So in the end, a de Rham cohomology class may be seen as a big collection of polynomials. In our discussion below, we do not need to bound the degrees of these polynomials but for any kind of effective implementation that would obviously be a good idea. See for example the work of Fabre giving an explicit complex calculating the Dolbeault cohomology [18]. 3.2. Computation of the Hodge integrals Suppose ϕ ∈ H Di R (X ) is an explicitly given algebraic de Rham cohomology class defined over Q, and suppose η ∈ Hi (X, Q) is an explicitly given cohomology class as described previously. We briefly explain (without R filling in too many details) how the integral η ϕ can be computationally approximated. The notion of constructive Hodge decomposition is discussed in Mitrea and Mitrea [40]. On each of the open balls covering our “puffed-up neighborhood” P of X , we may assume that there is an approximate linearization of the equations of X . Using this, we may define a version of Newton’s method to approximate a retract from P to X . Furthermore we may use C k piecewise real polynomial partitions of unity to extend the vector bundles of differential forms on X to C k piecewise real polynomial vector bundles defined on P, and apply the Bott–Tu method again using C k piecewise real polynomial partitions of unity. Thus, our element ϕ corresponds to an approximate differential form on P. On the other hand we can use the Newton-method retractions, glued together with a partition of unity, to map the simplicial complex H(X ) arbitrarily close to X itself, in C 1 norm. Then the integral may be calculated to a given precision, using interval arithmetic [19,20,26,34]. Proposition 3.2. Suppose ϕ ∈ H Di R (X ) is an explicit algebraic de Rham cohomology class defined over Q and suppose η ∈ Hi (X, Q) is an explicit homology class. Then we can compute a sequence of values αi (η, ϕ) ∈ Q such that Z αi (η, ϕ) − ϕ < 2−i . η

Proposition 3.3. Given X and η ∈ H2 p (X, Q) there is an algorithm which tries to show that η is not a Hodge class. 2p If it stops then it gives an algebraic de Rham cohomology class ϕ ∈ F p+1 H D R (X ) with coefficients in Q, and an i −i such that |αi (η, ϕ)| > 2 which shows that η is not a Hodge class. If η is not a Hodge class then the algorithm stops. 2p

Proof. Fix an affine open covering. The set of classes in F p+1 H D R (X ) is enumerable (actually the de Rham cohomology group should be computable as a finite-dimensional Q vector space but we are not using that). The algorithm is to enumerate the classes ϕ, and compute the αi (η, ϕ) Rfor bigger and bigger i, stopping if we obtain |αi (η, ϕ)| > 2−i . If η is not a Hodge class, there is some ϕ such that η ϕ 6= 0. As pointed out above we may assume R ϕ defined over Q, and there is some i such that | η ϕ| ≥ 21−i . For these values we will have |αi (η, ϕ)| > 2−i . Therefore, if η is not a Hodge class, the algorithm will eventually find the proof (ϕ, i).  An alternative approach would be to compute an approximation to the harmonic representative for a cohomology class, and see whether it is of type ( p, p). This could be done with a finite-element method, see [15,21,24,27] for example; or by approximating the heat equation, or maybe also by using geometrical approximations akin to [16]. 3.3. The full Hodge algorithm Suppose we have an explicit smooth projective variety X defined over Q, and suppose η ∈ H2 p (X, Q). Run in parallel (by alternating steps) the algorithm of Proposition 2.8 to try to show that η is an algebraic cycle, and the

136

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

algorithm of Proposition 3.3 to try to show that η is not a Hodge cycle. Call this combined algorithm, the full Hodge algorithm for (X, η). If it stops, it either shows that η is an algebraic cycle (2.8), or else gives a proof that η is not an algebraic cycle by proving that it is not a Hodge cycle (3.3). Theorem 3.4. The Hodge conjecture is equivalent to saying that the Hodge algorithm always stops. Proof. The Hodge conjecture says that any class η is either an algebraic cycle, or not a Hodge cycle. This is clearly equivalent to the statement that one of the two component algorithms stops.  Corollary 3.5. A proof of the Hodge conjecture would constitute a proof that the Hodge algorithm always stops. Corollary 3.6. The Hodge conjecture would imply a positive answer to Question 2.9, decidability of whether or not a given class is an algebraic cycle. We can state the analogue of Question 2.9 for Hodge cycles. Question 3.7. Is it decidable whether η ∈ H2 p (X, Q) is a Hodge cycle? The Hodge conjecture would also give a positive answer to this question. We can introduce variants on both questions, by starting with the Poincar´e dual of an algebraic de Rham cohomology class defined over Q rather than with a rational homology class. The Hodge conjecture also would give positive answers to those questions. If X doesn’t have coefficients in Q then we need to say more. For this we should introduce a notion of decidable subring Υ ⊂ C containing an enumerable list of elements on which the ring operations and equality are decidable, and allowing arbitrary approximation of the complex values of elements. Then we could use the same approximation techniques as before and get algorithms as in 2.8 and 3.3. 3.4. The Lefschetz (1, 1)-theorem The Lefschetz (1, 1)-theorem says that the Hodge conjecture is true for classes of complex codimension 1, and the Lefschetz decomposition gives the same for complex dimension 1. Corollary 3.8 (Lefschetz (1, 1)). Suppose X has dimension d. For classes η ∈ H2d−2 (X, Q) and η ∈ H2 (X, Q) the Hodge algorithm stops. Proof. The Lefschetz (1,1) theorem and its Poincar´e dual version give the Hodge conjecture in these cases.  In particular, Questions 2.9 and 3.7 have positive answers for η ∈ H2d−2 (X, Q) and η ∈ H2 (X, Q). This already seems like a somewhat remarkable logical situation. The known proofs of the Lefschetz (1, 1)-theorem (3.8) are transcendental, indeed the statement on the surface of it seems to be transcendental. However, the algorithm is combinatorially well-defined, so its stopping is a combinatorial question and we can ask whether there are combinatorial proofs, for example are there proofs valid in an intuitionist logical setting? Similarly, is there an effective bound on how long the algorithm runs before stopping? A related question is to bound the degree of the effective divisors whose difference gives our divisor class. This should be related to Seshadri constants and things like that [33]. 3.5. Absolute Hodge cycles Deligne defined a notion of absolute Hodge cycle in [13]. This is a cycle in the algebraic de Rham cohomology which is Hodge for every embedding Q ,→ C. This notion seems very close to the notion of decidability which we are considering here. Deligne proved that all Hodge cycles are absolute Hodge on abelian varieties, and used his techniques to give some computations of periods [11–13]. More recently see Voisin [55]. We conjecture that the same techniques, if investigated more closely, would give decidability of the Hodge cycle question for abelian varieties. Conjecture 3.9. The same techniques used by Deligne to prove the absolute Hodge property, give decidability of the Hodge cycle Question 3.7 or its variant in algebraic de Rham cohomology, for cycles on abelian varieties.

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

137

4. Things to compute Now here is the point we would like to emphasize: proceeding from these purely theoretical considerations of decidability, it is natural to ask whether the algorithm could in any sense be put into effect. In the general setting of an arbitrary variety, this is likely to lead to exponentially long computations or worse. However, it does suggest that it might be good to use the rapidly evolving technology in computer algebraic geometry, to carry out a more limited search for algebraic cycles. The hope is that such a search could help in our understanding of the problems surrounding algebraic cycles. For example, if a computer search turns up some interesting new algebraic cycles, then we might be able to see a pattern which could lead to a theoretical construction. On the other hand, if an extensive and reasonable-looking computer search failed to turn up any cycles representing a plausible Hodge class, then that would be evidence against the Hodge conjecture. One of the least computationally effective parts of our procedure is calculating the Hodge integrals trying to show that a given class is not Hodge. Also, if we choose a cohomology class “at random” then there is not that much reason to expect it to be Hodge. It would seem more reasonable to try to start with classes which are known for some abstract reason to be Hodge classes, and just look for the algebraic representatives. For these reasons we narrow things down in the following subsections. Of course, it would be better to avoid cases where the Hodge conjecture is already known, see [35] or type “Hodge conjecture” into a search engine. 4.1. Standard classes There are numerous well-known reasons for a class to be a Hodge class. The main examples come from Grothendieck’s standard conjectures [22,31]. These concern classes such as the Kunneth projectors, Lefschetz operators etc. which for theoretical reasons are automatically Hodge classes. We state some examples of problems which one could try to attack. The K¨unneth formula for the homology of a product yields integral classes which, because of the natural properties of the Hodge decomposition, are automatically of type ( p, p). For any X and Y , this gives classes on X × Y × X × Y . Since we would like to concentrate on dimensions as small as possible, it is better to start by looking at a related collection of classes on X × X . These are correspondences which, in the K¨unneth decomposition and via Poincar´e duality, give the identity on one of the Hi . Problem 4.1. For a specific smooth d-dimensional projective variety X of small degree, find algebraic cycles representing the K¨unneth projector classes ki ∈ H2d (X × X, Q) which cohomologically represent the projection H· → Hi . Another source of natural cohomology classes is the Lefschetz operator. Cup product with the hyperplane class ω has an adjoint operator L, which is rational and of type ( p, p). Viewed as a cohomological correspondence, it is a Hodge class on X × X , known to be algebraic only in some special cases [23,36,52]. Once again, to get a lowdimensional situation, it is probably a good idea to look at a related statement for classes on X itself, also for this statement we do not need to construct the adjoint L. Problem 4.2. For a specific smooth d-dimensional projective variety X of small degree, let ω ∈ H2d−2 (X, Q) be the homology class of the hyperplane section. Suppose we have an algebraic cycle η ∈ H2q (X, Q) for q < d, find an algebraic cycle ξ ∈ H2d−q (X, Q) such that ξ ∩ ωd−q = η. In order for these problems to be interesting, we have to start with a variety X such that the homology H· (X, Q) is sufficiently interesting. For instance it should not all be concentrated essentially in the middle dimension or come from some obvious geometric construction such as intersection of hyperplane sections or exceptional loci of birational transformations. Otherwise the trend is that the K¨unneth projectors can be shown to be algebraic, as has already been done in some cases by Saito [52], Bloch–Esnault [3]. It seems already to be an interesting and nontrivial computational problem to find small and easy varieties with homology which is sufficiently interesting, for example spread out in many degrees, that the search for K¨unneth projectors would not have an imaginable theoretical answer. 4.2. Classes determined by monodromy There can be other reasons for obtaining Hodge classes, for instance due to monodromy. If X → Z is a smooth projective family over a quasiprojective base, then the homology groups of the fibers fit into a Q-variation of Hodge

138

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

structures Hi (X/Z ) of weight i over Z . This data includes the monodromy representation of π1 (Z , z 0 ) on the fiber Hi (X z 0 , Q) as well as the Hodge decompositions of the fibers Hi (X z , Q) ⊗Q C for all z ∈ Z . A classical result in Hodge theory is that the subspace of invariants is a sub-Hodge structure of weight i in each fiber. The highest wedge power of the subspace of invariants is therefore a rank one Hodge structure which must consist of Hodge cycles. To say this more precisely, let Γ (Z , Hi (X/Z )) denote the space of global sections of Vi over Z . For each z ∈ Z we have Γ (Z , Vi ) → Vi (z) = Hi (X z , Q). Suppose Γ (Z , Hi (X/Z )) has dimension k. Then for any z ∈ Z ,   ^ k O ik ∼ k Q Γ (Z , Hi (X/Z )) ,→ Hi (X z , Q) = Hik (X zk , Q) = 2 is a morphism of Hodge structures, so the image of a generator is a Hodge cycle in Hik (X zk , Q). So we can try to look for an explicit algebraic cycle representing it. This situation can arise when we have two families. Suppose X, Y → Z are families, and suppose that the monodromy representations Hi (X/Z ) and Hi (Y /Z ) contain a common representation V of π1 (Z , z) as a factor. Then H2i (X × Z Y /Z ) contains a nontrivial family of fixed vectors, and the above discussion applies. If k is the rank of the space of fixed vectors (which depends on the structure of the monodromy representations) then we obtain a family of algebraic cycles in (X z × Yz )k for every z ∈ Z . One would like to have an example where k is small. This kind of consideration goes into the construction of Hodge cycles in a recent paper of Voisin [55]. In particular, one could search for explicit algebraic representatives for the cycles constructed in Example 3.4 of [55]. One can ask more generally for other examples of this kind. See [43]. Katz’s algorithm for rigid local systems [30,14,50], run in different ways to give different motivic expressions for the same local system, furnishes an interesting class of examples. 4.3. Ways of finding algebraic cycles A brute-force enumeration of all possible smooth varieties of Y mapping to our given variety of X is probably not the most efficient way to look for algebraic cycles on X . There are formulations which should be more convenient and reduce the complexity of the problem. We mention here Thomas’s theorem [54]. It says that if a class η ∈ H2 p (X, Q) is represented by an algebraic cycle with p ≤ d − 2, then there is a hypersurface H ⊂ X in a multiple of the standard hyperplane class, such that H has only isolated double points as singularities, such that if H 0 denotes the variety resolved by blowing up the double points then η is in the image of H2 p (H 0 , Q) → H2 p (X, Q). On the other hand, if η is a Hodge class and if it comes from a singular hypersurface, then the Hodge conjecture in smaller dimension would give an algebraic cycle for it. For example in codimension 2 this is sufficient because on H 0 the problem turns into the Lefschetz (1, 1) theorem. Another way we could get an explicit construction of algebraic cycles would be to look at complete intersections of hyperplane sections, then take an irreducible component—which is essentially a problem of factorization [9]. This construction will give all possible subvarieties. From these remarks, it is clear that we need to master computations such as finding and resolving the singularities of a singular hypersurface, or computing the irreducible components of a complete intersection. 4.4. Conclusion A general review of the relationship between decidability or constructability, and the Hodge conjecture, leads naturally to the question of doing a computational search for algebraic cycles representing Hodge classes. It also turns out to be a nontrivial question to first construct the varieties which provide interesting examples where this computation should be done. These computational problems point to the importance of specific techniques in constructive algebraic geometry. References [1] S. Basu, Computing the top Betti numbers of semi-algebraic sets defined by quadratic inequalities in polynomial time. Preprint math.AG/0603262.

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

139

[2] S. Basu, R. Pollack, M.-F. Roy, Algorithms in Real Algebraic Geometry, second edition, in: Algorithms and Computation in Mathematics, vol. 10, Springer-Verlag, 2006. [3] S. Bloch, H. Esnault, K¨unneth projectors for open varieties. Preprint math.AG/0502447. [4] J. Bochnak, M. Coste, M.-F. Roy, Real Algebraic Geometry, in: Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 36, Springer-Verlag, 1998. [5] R. Bott, L. Tu, Differential Forms in Algebraic Topology, in: Graduate Texts in Mathematics, vol. 82, Springer, 1982. [6] D. Brownawell, Bounds for the degrees in the Nullstellensatz, Ann. of Math. 126 (1987) 577–591. [7] G. Carlsson, V. de Silva, Topological approximation by small simplicial complexes. Preprint available at: http://math.stanford.edu/∼silva/ public/papers/small.pdf. [8] G. Carlsson, A. Zomorodian, Computing persistent homology, Discrete Comput. Geom. 33 (2005) 249–274. [9] G. Ch`eze, A. Galligo, From an approximate to an exact absolute polynomial factorization, J. Symbolic Comput. 41 (2006) 682–696. [10] Martin Davis, Hilary Putnam, Julia Robinson, The decision problem for exponential diophantine equations, Ann. of Math. (2) 74 (1961) 425–436. [11] P. Deligne, Travaux de Shimura, in: S´eminaire Bourbaki (1970–71) 389, in: Lecture Notes in Math., vol. 244, Springer, 1971, pp. 123–165. [12] P. Deligne, Vari´et´es de Shimura: interpr´etation modulaire, et techniques de construction de mod`eles canoniques, in: Automorphic Forms, Representations and L-Functions (Corvallis, 1977), in: AMS Proc. Sympos. Pure Math., vol. 23, 1979, pp. 247–289. [13] P. Deligne, J.S. Milne, A. Ogus, K. Shih, Hodge Cycles, Motives, and Shimura Varieties, in: Lecture Notes in Mathematics, vol. 900, 1982. [14] M. Dettweiler, S. Reiter, An algorithm of Katz and its application to the inverse Galois problem, in: Algorithmic Methods in Galois Theory, J. Symbolic Comput. 30 (2000) 761–798. [15] J. Dodziuk, Finite-difference approach to the Hodge theory of harmonic forms, Amer. J. Math. 98 (1976) 79–104. [16] S. Donaldson, Scalar curvature and projective embeddings. I, J. Differential Geom. 59 (2001) 479–522. [17] H. Edelsbrunner, D. Letscher, A. Zomorodian, Topological persistence and simplification, in: Discrete and Computational Geometry and Graph Drawing (Columbia, SC, 2001), Discrete Comput. Geom. 28 (2002) 511–533. [18] B. Fabre, Locally residual currents and Dolbeault cohomology on projective manifolds, Bull. Sci. Math. 130 (2006) 553–564. [19] C. Fefferman, The N -body problem in quantum mechanics, in: Frontiers of the Mathematical Sciences: 1985 (New York, 1985), Comm. Pure Appl. Math. 39 (1986) S67–S109. [20] C. Fefferman, R. de la Llave, Relativistic stability of matter, I, Rev. Mat. Iberoamericana 2 (1986) 119–213. [21] J. Friedman, Computing Betti numbers via combinatorial Laplacians, Algorithmica 21 (1998) 331–346. [22] A. Grothendieck, Standard conjectures on algebraic cycles, in: Algebraic Geometry (Internat. Colloq., Tata Inst. Fund. Res., Bombay, 1968), Oxford Univ. Press, 1969, pp. 193–199. [23] B. Gordon, M. Hanamura, J. Murre, Absolute Chow-K¨unneth projectors for modular varieties, J. Reine Angew. Math. 580 (2005) 139–155. [24] P. Hanlon, Hodge structures on posets, Proc. Amer. Math. Soc. 134 (2006) 1857–1867. [25] T. Hales, A proof of the Kepler conjecture, Ann. of Math. 162 (2005) 1065–1185. [26] J. Hass, R. Schlafly, Double bubbles minimize, Ann. of Math. 151 (2000) 459–515. [27] R. Hiptmair, Discrete Hodge operators, Numer. Math. 90 (2001) 265–289. [28] M. Hirsch, S. Smale, On algorithms for solving f (x) = 0, Comm. Pure Appl. Math. 32 (1979) 281–313. [29] E. Hrushovski, Computing the Galois group of a linear differential equation, in: Differential Galois Theory (Bkhedlewo, 2001), in: Banach Center Publ., vol. 58, Polish Acad. Sci., Warsaw, 2002, pp. 97–138. [30] N. Katz, Rigid Local Systems, in: Annals of Mathematics Studies, vol. 139, Princeton University Press, 1996. [31] S. Kleiman, Algebraic cycles and the Weil conjectures, in: Dix esposs sur la cohomologie des sch´emas, North-Holland, 1968, pp. 359–386. [32] S. Lang, Diophantine Geometry, in: Interscience Tracts in Pure and Applied Mathematics, vol. 11, Interscience Publishers, John Wiley & Sons, 1962. [33] R. Lazarsfeld, Lengths of periods and Seshadri constants of abelian varieties, Math. Res. Lett. 3 (1996) 439–447. [34] R. de la Llave, Computer assisted proofs of stability of matter, in: Computer Aided Proofs in Analysis (Cincinnati, OH, 1989), in: IMA Vol. Math. Appl., vol. 28, Springer, 1991, pp. 116–126. [35] J. Lewis, A Survey of the Hodge Conjecture, second edition, in: CRM Monograph Series, vol. 10, AMS, 1999, Appendix B by B. Brent Gordon. [36] D. Lieberman, Numerical and homological equivalence of algebraic cycles on Hodge manifolds, Amer. J. Math. 90 (1968) 366–374. [37] A. Macintyre, Model theory: geometrical and set-theoretic aspects and prospects, in: New Programs and Open Problems in the Foundation of Mathematics (Paris, 2000), Bull. Symbolic Logic 9 (2003) 197–212. [38] B. Mazur, Questions of decidability and undecidability in number theory, J. Symbolic Logic 59 (1994) 353–371. [39] J. Matijaseviˇc, The Diophantineness of enumerable sets, Dokl. Akad. Nauk SSSR 191 (1970) 279–282 (in Russian). [40] D. Mitrea, M. Mitrea, Finite energy solutions of Maxwell’s equations and constructive Hodge decompositions on nonsmooth Riemannian manifolds, J. Funct. Anal. 190 (2002) 339–417. [41] L. Moret-Bailly, Elliptic curves and Hilbert’s tenth problem for algebraic function fields over real and p-adic fields, J. Reine Angew. Math. 587 (2005) 77–143. [42] B. Mourrain, F. Rouillier, M.-F. Roy, The Bernstein basis and real root isolation, in: Combinatorial and Computational Geometry, in: Math. Sci. Res. Inst. Publ., vol. 52, Cambridge Univ. Press, 2005, pp. 459–478. [43] H. Movasati, S. Reiter, Hypergeometric series and Hodge cycles of four dimensional cubic hypersurfaces. Preprint math.AG/0507436. [44] A. Nabutovsky, Isotopies and nonrecursive functions in real algebraic geometry, in: Real Analytic and Algebraic Geometry (Trento, 1988), in: Lecture Notes in Math., vol. 1420, 1990, pp. 194–205.

140

C. Simpson / Theoretical Computer Science 392 (2008) 128–140

[45] A. Nabutovsky, Fundamental group and contractible closed geodesics, Comm. Pure Appl. Math. 49 (12) (1996) 1257–1270 (English summary). [46] A. Nabutovsky, S. Weinberger, Algorithmic unsolvability of the triviality problem for multidimensional knots, Comment. Math. Helv. 71 (1996) 426–434. [47] P. Niyogi, S. Smale, S. Weinberger, Finding the homology of submanifolds with high confidence from random samples. Preprint homolog of submanifolds from data.pdf. Available at: http://ttic.uchicago.edu/∼smale/papers/. [48] J. Parker, Talk, Max-Planck Institut f¨ur Mathematik, September 2006. [49] G. Prasad, S. Yeung, Fake projective planes. Preprint math.AG/0512115. [50] D. Roberts, Rigid Jordan tuples. Preprint available at: http://cda.morris.umn.edu/∼roberts/. [51] J. Robinson, Unsolvable diophantine problems, Proc. Amer. Math. Soc. 22 (1969) 534–538. [52] M. Saito, Chow-Kunneth decomposition for varieties with low cohomological level. Preprint math.AG/0604254. [53] S. Smale, Algorithms for solving equations, in: Proceedings of the International Congress of Mathematicians, vol. 1, 2, Berkeley, CA, 1986, Amer. Math. Soc., 1987, pp. 172–195. [54] R. Thomas, Nodes and the Hodge conjecture, J. Algebraic Geom. 14 (2005) 177–185. [55] C. Voisin, Hodge loci and absolute Hodge classes. Preprint math.AG/0605766. [56] Y. Yomdin, Global bounds for the Betti numbers of regular fibers of differentiable mappings, Topology 24 (1985) 145–152.